Using RL, How to train multi-agents such that each agent will navigate from its initial position to goal position avoiding collisions?

Let's assume there are a set of agents that are spread into 3d cartesian space. A trajectory should be generated for each agent such that if an agent would follow its trajectory while heading to the goal waypoint, no collision would happen with other agents. Any guidance to solve such a task would be highly appreciated

Answer 1

[Emmanouil Tzorakoleftherakis]

It's possible that the scenario you described can be solved by training a single agent, and then "deploying" that trained agent to all uavs/uuvs in your fleet. That would make the problem easier and less expensive to train. For a 2D example, take a look at this.

[Steve Jessy]

@Emmanouil Tzorakoleftherakis

Thanks, for your comment.

What you mentioned is correct how that would make the problem easier and less expensive to train. But how then each agent will avoid the crash with other agents while navigating?

Also, could you please assist me in how can I apply the scenario taking into consideration that agent doesn't have sensor detects obstacles, the navigation is based on odom data provided by the agent

[Emmanouil Tzorakoleftherakis]

I think it's a matter of what inputs you provide to the policy and the coordinate system you use (although I was thinking the scenario where each agent has its own sensors). If you only use odometry data from all agents, I guess you could transform it to distance from each nearby agent (include heading/bearing probably) and feed all this info into the policy.

[Steve Jessy]

@Emmanouil Tzorakoleftherakis

The coordinate system in which the agents are acting is 2D cartesian coordinate system. Yes I can access the distance from an agent to all the other agents in the space. I'd like to kindly ask you if you can provide an example/code in which the multi-agent system is trained based on odom data